گروک 4.5

SpaceXAI's intelligent coding model for agentic software, engineering and workflow tasks, trained on new datasets spanning science, engineering and mathematics. It handles everything from difficult Rust and C or C++ work to building end-to-end applications from prompt to production, and is highly proficient at well-designed, functional apps. The lab reports roughly twice the token efficiency of comparable leading models; reasoning is always on.

xAI trained Grok 4.5 alongside Cursor and aimed it squarely at real engineering work. The run used tens of thousands of NVIDIA GB300 GPUs with stability techniques built for training at that scale, and the lab says it spent as much effort on the data as on the volume: deduplication, quality scoring and domain-focused selection to keep the mixture both broad and high-signal. Reinforcement learning then covered hundreds of thousands of tasks centred on multi-step software engineering and other technical work, graded automatically and by model, on a stack built for asynchronous training so agentic rollouts can run for hours while learning continues across the cluster.

The efficiency claim is the one xAI leads with, and it is unusually specific. The lab reports Grok 4.5 resolving SWE Bench Pro tasks with 15,954 output tokens on average against 67,020 for Opus 4.8 at maximum effort - about 4.2 times fewer - while being served at fast-model speeds of 80 tokens per second. Roughly twice the token efficiency of comparable leading models, solving tasks in under half the number of steps, is the trade xAI says it is offering: not the highest peak score, but the most intelligence per unit of time and cost.

The five evaluations xAI published bear that reading out. It leads one of them outright, reporting 29.0% on SWE Marathon against 26.0% for Opus 4.8 at max effort and 24.0% for Fable at max. On the other four it is close but behind: 83.3% on Terminal Bench 2.1 against a leading 84.3%, 62.0% on DeepSWE 1.0 against 66.1%, 53% on DeepSWE 1.1, and 64.7% on SWE Bench Pro against 80.4%.

Away from code, xAI made Grok 4.5 the default model in its build product and pushed it at document work. The lab describes it building complex Excel models that research the web, use formulas across multiple sheets and leave notes behind for whoever opens the file next; assembling PowerPoint diagrams out of native shapes rather than pasted pictures; and drafting clear prose in Word. Plugins for Word, PowerPoint, Excel and Outlook exist so that work can happen where the documents already live.

Read the table honestly and the case for Grok 4.5 is speed and cost at near-frontier quality, not a top score: it trails the strongest models on three of the five evaluations its own launch post publishes. Reasoning cannot be turned off, and the effort dial stops at high - a request asking for the xhigh setting is served as high, because that level arrived with the following generation. xAI's own current guidance points developers at its newer flagship for both code and chat, which makes this the model to choose deliberately, for its token economics, rather than by default.